U.S. patent application number 11/040632 was filed with the patent office on 2006-07-27 for predictive artificial intelligence and pedagogical agent modeling in the cognitive imprinting of knowledge and skill domains.
Invention is credited to Dean Gordon Arrasmith, Deme Michael Clainos, T. Peter Rowe.
Application Number | 20060166174 11/040632 |
Document ID | / |
Family ID | 36697235 |
Filed Date | 2006-07-27 |
United States Patent
Application |
20060166174 |
Kind Code |
A1 |
Rowe; T. Peter ; et
al. |
July 27, 2006 |
Predictive artificial intelligence and pedagogical agent modeling
in the cognitive imprinting of knowledge and skill domains
Abstract
System and methods for predicting and dynamically adapting the
most appropriate content and teaching strategies that aid
individual student learning. System and methods are based on a
cognitive model that integrates new information with what the
student already knows. A program of study is predicted by the
unique cognitive needs of the individual student correlated with
aggregated student data history using an Artificial Intelligence
Engine (AI Engine). Said system and methods then dynamically adapt
the initial cognitive model to the student's ongoing progress using
personalized software Agents. Said system and methods include a
computer network that incorporates a server-side AI Engine and a
collection of client-side software Agents embodied as animated
characters. The program connects new information to prior knowledge
and then strengthens these connections through dedicated learning
Activities, customized to the student, to ensure that effective,
and real, learning occurs.
Inventors: |
Rowe; T. Peter; (Portland,
OR) ; Arrasmith; Dean Gordon; (Aurora, OR) ;
Clainos; Deme Michael; (Lake Oswega, OR) |
Correspondence
Address: |
Deme M. Clainos;StudyDog
9720 SW Nimbus Ave.
Beaverton
OR
97008
US
|
Family ID: |
36697235 |
Appl. No.: |
11/040632 |
Filed: |
January 21, 2005 |
Current U.S.
Class: |
434/236 |
Current CPC
Class: |
G09B 5/06 20130101 |
Class at
Publication: |
434/236 |
International
Class: |
G09B 17/00 20060101
G09B017/00; G09B 19/00 20060101 G09B019/00 |
Claims
1. A method of ensuring that students learn from instruction,
comprising the steps of: acquiring cognitive student data into a
computer; storing cognitive student data models in a computer;
automatically creating a customized program of content activities
based on a student cognitive model; and automatically adjusting a
customized program of activities based on a changing student
cognitive model as said student progresses through the program
2. The method of claim 1, wherein acquiring cognitive student data
step is the input of prior knowledge from familiar sources and
directly from student
3. The method of claim 1, wherein the computer has an Artificial
Intelligence Engine to automatically create a customized program of
activities step
4. The method of claim 1, wherein the automatic adjustment of the
customized program of activities step is implemented by software
agents
5. The method of claim 1, wherein a customized program of
activities is managed by a helper agent and a plurality of content
agents
6. The method of claim 3, wherein the automatic creating of a
customized program step further comprises the step of pattern
matching the target students individual cognitive model with the
historically stored cognitive model of all previous students using
an Artificial Intelligence Engine
7. The method of claim 5, wherein the agent further comprises a
helper agent to guide and encourage the student, and multiple
content agents to present instructional material.
8. The method of claim 5, wherein automatic adjusting of the
student cognitive model is implemented by the helper agent and the
content agents monitoring student responses to customized program
of activities.
9. The method of claim 8 wherein the helper agent and the multitude
of contents agents communicate with the student via computer
response, voice recognition and speech.
10. A computer-implemented learning system, comprising: a server
computer and a plurality of client computers on a network means for
displaying a prior knowledge questionnaire and test to a
parent/guardian/teacher and student on a client computer, and
storing said results in a student cognitive model dataset means for
storing a student cognitive model in a server computer means for
comparing a multitude of stored student cognitive models with a new
student cognitive model using an Artificial Intelligence Engine for
the purpose of identifying pattern matches of past successful
activity programs means for downloading and implementing software
helper and content agents on a client computer means for downloaded
and visually presenting content activities to a student on a client
computer means for student to interact with content activities and
to have said interactions stored means for communications between a
server AI Engine and client software agents
11. The apparatus of claim 10, further comprising a means to modify
a student cognitive model according to the student's results
12. The apparatus of claim 10, further comprising a means to hear
student directions and questions using voice recognition and a
means to instruct the student using speech.
13. The apparatus of claim 10, further comprising means of
providing parents/guardians/teachers with reports based on student
interactions with content activities.
14. The apparatus of claim 10, further comprising a mechanism to
allow students to accumulate rewards in the form of points based on
their results of mastering content activities.
Description
CROSS REFERENCE TO RELATED PULICATION
[0001] This application claims the benefit of the U.S. Provisional
Patent Application 60/538,030 with the filing date of Jan. 22,
2004. TABLE-US-00001 5761649 June, 1998 Hill 5974446 October, 1999
Sonnenreich et al. 5978648 November, 1999 George et al. 6035283
March, 2000 Rofrano. 6155840 December, 2000 Sallette 6201948 March,
2001 Cook et al. 6237035 May, 2001 Himmel et al. 6321209 November,
2001 Pasquali. 6343329 January, 2002 Landgraf et al. 6356284 March,
2002 Manduley et al. 6427063 July, 2002 Cook et al. 6470171
October, 2002 Helmick et al. 6,845,229 Jan. 18, 2005 Educational
instruction system
BACKGROUND OF THE INVENTION
[0002] The use of technology in learning is still in its infancy
but it has the potential to significantly impact our educational
system in a positive way. Thus far, instructional technology has
been mostly focused on visually presenting content organized in a
mostly static way. This is understandable since a significant asset
of the computer is that it is a medium-rich environment with
features such as sound, movies, text, speech recognition,
handwriting analysis, networked community environments, and
multiplayer functionality that allow for the creation of a dynamic
and exciting environment for the student. The difficulty has been
harnessing the decision-making power of the computer to provide
more effective learning environments. The realities of how these
features are integrated into present-day computer-based
instructional systems is generally simplistic: if a student incurs
errors above a certain threshold in a certain area of study then
they are deemed as lacking in that area and additional or
alterative material is presented. Likewise if the student does
exceptionally well in a specific activity or test then alternative,
more difficult, material is presented. In general these systems
start with each student as a new entity and address the
presentation of instructional materials in a pragmatic way. In some
cases these systems classify the relatively short-term progression
history of the student and organize the sequence of instruction
presentation on this basis. In other cases the program "adapts" to
the child but typically this is a solution incorporating "fixed"
content using simple branching logic. These are pragmatic solutions
that make effective use of the computers visual and audio
capabilities and concentrate more on presenting their content to
the child, rather than letting the child's current state of
intellect influence how and what is drawn to them.
[0003] Cognitive Science focuses on correlating instructional
strategies and content with how the brain really works. Using a
Cognitive Model as a basis for learning provides strategies more
aligned with how receptive the brain is to receiving new
information and how that information can be learned so that it is
retained, and recalled later in a meaningful and useful way.
[0004] For example, learning to read is not an isolated skill and
is much more involved than passive "decoding" skills. The cognitive
model of reading is about the reader bringing their world
experiences to bear on integrating new information with what they
already know. Reading is not passive, it is active and the context
depends very much on who the reader is. The same passage read by
one reader may have a totally different meaning to another. Even
very young readers engage many thought processes when reading such
as predicting, categorizing, and making unexpected connections. At
the same time they use strategies for comprehending words,
sentences, and segments of text. Readers also make use of
non-verbal clues from pictures, color, typography, and layout. In
cognitive terms readers are active, selective, and strategic, they
understand how and what they read in terms of what they already
know, and they use many different thought processes in its
execution. Many of today's computer-based learning systems are not
interesting to the student because they do not operate within the
same context, or world-view, in which the student resides. By
relating to the cognitive needs of the student, programs can be
created that align with their needs. Rather than force a specific
top-down curriculum on the student, the program provides the
student with what their learning needs require. This makes the
program more effective in helping them learn, while providing a
more enjoyable learning environment for the student.
BRIEF SUMMARY OF THE INVENTION
[0005] The present invention is based around a "Cognitive Model" of
each student. This cognitive model reflects the child's preferences
and what they already know, and is initially built with data from
external sources such as parental input, teacher input, student
achievements, student questionnaires and testing. Some examples
might include the sports or activities the student likes to play,
the movies they may have seen, the stories they are familiar with,
the hobbies they like most, the people they know.
[0006] The present invention incorporates a neural-net based
Artificial Intelligence Engine (AI Engine) that discovers patterns
between the cognitive model of a new student and the collective
cognitive models of a population of previous students. In this way
the AI Engine can initially predict the most effective program of
study for a child based on its past experience with other students.
It is this prediction data that is used to initially populate areas
of the individual Student Cognitive Model, and to assign an initial
program of study to new students. As more students use the system
more data is available for the AI Engine to make more accurate
predictions about new students.
[0007] In the present invention, as a student progresses through a
program of study, wherein responses and results can be measured, a
series of Intelligent Pedagogical Software Agents or "Agents" are
assigned to, and learn more about, each individual student. The
Agents fine-tune and adapt to the current, and ongoing, cognitive
state of the student to provide real-time alignment to the best
ways that skills are imprinted. In this way a distributed system of
intelligent components is used to create and maintain a "virtual"
cognitive model of each student. This learning environment begins
with the best possible predicted course of study for each student
based on his or her cognitive model and then learns how to
fine-tune that environment to deliver a uniquely personalized
program. Each program is based on the current and ongoing cognitive
model of the student, and of the cognitive goals of the program.
This results in learning by the most efficient and effective means
for each student.
[0008] In the present invention the factors governing this learning
process are based on the unique cognitive makeup and cognitive
outcome requirements of each student. The rules of the system are
directed at the highest level by the cognitive needs of the student
and the cognitive goals of the program.
[0009] In the present invention, this system, as described, begins
with the background data of the student's prior knowledge. Further
information from the results of student tests is also incorporated.
A predictive Artificial Intelligence Engine (AI Engine) then
initializes a set of data in a cognitive model for the student.
This model is patterned with previous aggregate student cognitive
model data to identify an initial customized program of study for
the individual student.
[0010] In the present invention, after the initial program of study
is provided to the student, a set of software Agents refine and
validate the predictions and further personalize the delivery of
instruction for each student based on their unique current and
ongoing cognitive state, and the goals of the program. The AI
Engine thereby initially predicts data components of the cognitive
model before the student begins a course of study while the Agents
then dynamically adapt these predictions as new information is
gained. The Agents negotiate to individualize the instructional
program for each student even further by learning and adapting to
how each student best responds to delivered instruction. This
intelligence is returned to the AI Engine's collective data to
allow for more accurate initial predictions for new students in the
future. The Agents operate in real-time on behalf of each
individual student and continually learn to identify and acquire
the most effective, accurate, and up-to-date instructional material
for that student based on their changing cognitive model. The
entire system creates a contextual environment that maps to the
cognitive state of the student thereby offering content and
teaching strategies that are aligned with the student's ability to
learn new information.
BRIEF DESCRIPTIONS OF THE DRAWINGS
[0011] FIG. 1 shows the components of a research based cognitive
learning model and how new information is integrated with prior
knowledge, and how though learning the new information is
integrated as part of the knowledge pool, or modifies what is
already in the pool.
[0012] FIG. 2 shows the implementation of the Student Model that
incorporates the Cognitive Model of the Students Prior Knowledge
Network, the Assessment Results, the Lessons Task Models/Results,
and the Student Concept Map.
[0013] FIG. 3 shows the external relationships that can impact the
learning state of the Student Cognitive Model. These consist of
Prior Knowledge, Culture, the AI Engine, the Predictive Classifier,
Parents & Teachers, and the Agents.
[0014] FIG. 4 shows the architecture of the Agents interfacing with
the Learning Environment. This interface is carried out through a
single unit called the Agent Control Module that in turn interfaces
with the Domain Knowledge, the Cognitive Model, the Pedagogical
Session Model, and the Agent Appearance and User Interface (UI)
Module.
[0015] FIG. 5 shows the components of the Agent Behavioral Control
Description. These include the Pedagogical Goals, The Potential
Student Actions, The Agent Interventions, and the Student Concept
Database.
[0016] FIG. 6 shows an overall high-level architecture of the
system. In the top section this shows the Student Cognitive Model
and how the Lesson and Assessment Results along with the Concept
Map of cognitive goals, and the Student profile, all relate to
provide the state of the Student Cognitive Model. The bottom
section shows how the AI Engine aggregates information from a
population of Student Cognitive Models to provide predictive data
to the Student Model.
[0017] FIG. 7 shows the components of the Student Learning Profile
divided into three sections: Skill Inventory, Personal
Demographics, and Learning Preferences. These sections are further
divided to show the Skill Inventory as being made up, for example,
of the National Reading Panel Skills, Six-Traits for Writing, or
NCTM Ten Math Areas, the Personal Demographics being made up of
Student Age, Gender, Family Structure and Income, Special Needs,
and the Learning Preferences being made up of the Preferred
Learning Style(s) and Preferred Learning Environment
[0018] FIG. 8 shows the Skill Level Instructional Model. This
begins with a Placement Test followed by a loop of Instruction,
Modeling, and Practice that is tested with a Benchmark Test to
identify the skill Range and to then teach new skills and reinforce
prior skills.
[0019] FIG. 9 shows a framework of learning systems curriculum
consisting of specific instructional knowledge and skill areas
derived from the research literature. These areas describe, at a
high level, student proficiency targets, taking into account
age/grade and appropriate content.
[0020] FIG. 10 shows an implementation relation between a Server
computer, which provides the software which implements the system
and stores all the data, to multiple client computers that support
Agents used by students. While some components of the system
execute on the Client, all of these components initially reside on
the Server Computer, and are acquired by the client. The server
computer can service many clients over a network such as the
Internet, or an internal network such as a Local Area Network
(LAN).
Tables
[0021] Table 1 illustrates the manner in which each of the modules
of the Student Cognitive Model is impacted by the various data
driven and intelligent components of the learning system.
TABLE-US-00002 TABLE 1 How Components of the Student Cognitive
Model are impacted Student Lesson Task Assessment Student Cognitive
Model Models/Results Results Concept Map Parents & Enters
initial May examine May examine Teachers known through a through a
information reporting reporting interface. interface. Machine
Consumes known Consumes Consumes Learning portions of student
lesson results as assessment Engine profiles as training training
data to results as data to build build predictive training data to
predictive classifier. build predictive classifier classifier.
Predictive Uses known May predict May predict Classifier portions
of student lesson results. assessment profile as input to results.
predict unknown information about the student. Will predict some
initially unknown elements of student profile. Instructional
Records lesson May use to Records Agents task model and assist with
observed results and uses customization student them to shape of
instructional concepts and instructional level. certainty
customization values, and in each agent's uses these to individual
shape lesson. customization in each agent's Lesson. StudyDog Uses
Student Uses lesson Uses Uses student Agent Cognitive Model to
results in lesson assessment concept map in customize lesson
selection and results in lesson lesson plan. May also use
intervention to selection. selection. information in change lessons
student profile to midstream. guide the ways in which the Helper
Agent interacts with the student.
DETAILED DESCRIPTION OF THE INVENTION
Cognitive Learning Models
[0022] The present invention is aligned with modern theories about
how the brain learns. Instead of a strategy of "one size fits all"
classroom presentations and basic teaching techniques such as the
decoding component of reading, new information is introduced into a
context of what the student already knows. This "integration" of
new information with prior knowledge results in comprehension. For
example information about a particular game is incomprehensible to
someone lacking prerequisite knowledge about the game but is easily
comprehended by someone who knows the rules and strategies of
play.
[0023] However, comprehension does not mean learning has occurred.
For learning to occur the connection between new information and
prior knowledge must be strengthened by activities to the extent
that the new information becomes a part of, or modifying, the
existing knowledge. Therefore learning takes place when new
information becomes part of existing knowledge. Beyond learning is
that the new information now integrated into the existing knowledge
base is meaningful and useful. Knowledge is "meaningful" only after
it is richly interconnected with related knowledge. Knowledge is
"useful" only if you can access it under appropriate circumstances.
Meaningful knowledge is filed and cross-referenced with other
knowledge to which it is connected. Useful knowledge is filed and
cross-referenced so that you can find it when you need it.
[0024] Once new information is learned and the knowledge exists the
issue now becomes how can it be a retrieved and acted upon at a
later date. The brain does not appear to store data in a linear
manner; it appears to be a network of connections to relevant
information. Once one piece of information is accessed other
relevant pieces then become available. This information generally
isn't stored with perfect accuracy; the brain appears to be able to
reconstruct a good "interpretation" of the information it already
has. The brain can even "remember" things it never learned by
inferring from the information it already knows.
Cognitive Imprinting Of Knowledge And Skill Domains
[0025] The knowledge and skills associated with reading, writing,
and mathematics have existed for the past 4,000 to 5,000 years--a
very short time in terms of human evolution. Only within the past
few centuries have many people become literate in these areas of
thought and activity. While many human attributes, such as speech
and counting, have evolved much earlier and are now naturally
learned systems in the human brain, the skills and knowledge
associated with reading, writing, and mathematics have to be
specifically taught and learned. Without specific instruction and
practice, these domains will likely not be manifested in
humans.
[0026] Unlike innate skills, the human brain must be imprinted with
the knowledge and skills from these domains, and neurological
networks established and strengthened through repetitive practice
to achieve proficiency. Neuroscientists recognize that as these
skills and knowledge develop they involve several brain areas
associated with visual recognition, memory, language processing,
speech, semantic understanding, and higher-order processing. Some
neuroscientists believe in a deeper cognitive processing that forms
the meta-cognition that is required to proficiently employ these
learned domains. Several brain sites must act in harmony, in
neuro-networks, for proficiency in these knowledge and skill
domains. Some of these sites rely on secondary use of the cortical
areas of the brain to do jobs that they were not originally
intended to do, a phenomenon identified by Darwin as a conversion
of function in anatomical continuity.
[0027] In order for the brain to develop the new functional centers
and to form and strengthen the neural-networks between the centers,
knowledge and skill must be carefully sequenced and explicitly
taught to build a scaffolding to support increasingly higher-order
activities. Basic, prerequisite skills must be sufficiently
developed and made automatic to allow cognitive attention for
meta-cognitive activities such as higher-order comprehension,
control of language and meaning, and inferential mathematical
reasoning, for examples. Basic and automatic understanding of
letter and sound associations, writing mechanics, and number-word
associations are required to develop higher order knowledge and
skills in reading, writing, and mathematics, respectively.
[0028] The present invention learning system embodies these
concepts within the framework of a Student Cognitive Model using
artificial intelligence sub-systems and software Agents to deliver
instruction of uniquely sequenced and recursive knowledge and
skills domains.
[0029] The present invention uses technology to teach students how
to read based on the above cognitive model. The description below
describes this in detail.
[0030] FIG. 1 shows a cognitive learning model where the connection
of new information and prior knowledge lead to comprehension.
Activities that strengthen this connection are shown as
"Elaboration" by which real learning occurs as connections are
strengthened and new information is added to the prior knowledge
network. The present invention implements this model by initially
creating a Prior Knowledge database for each student. This
information, in combination with information inferred by the AI
Engine based on the history of past students, and the student test
and questionnaire results are is used to initially create a
customized course of study for the student
[0031] FIG. 2 shows the Student Cognitive Model for the present
invention learning system. This Student Cognitive Model represents
all the student information that the system can use to customize
the learning experience on a per-student basis. This Student
Cognitive Model is used, and contributed to by the AI Engine, the
Helper and Content Agents, input from a parent and teacher, and by
student knowledge information from other sources. The Student
Cognitive Model is comprised of a number of modules which we will
now describe:
[0032] The `Student Cognitive Model" is the implementation of the
students cognitive state and contains the initial student profile
as input by parents and teachers, the student's learning style
information, and a knowledge base of any other general information
where the system might learn about what the student already
knows.
[0033] The `Lesson Task Model/Results` module represents
essentially what the present invention learning system tracks on a
per-lesson basis--where the student has progressed to in a
particular lesson, difficulties they have had, time spent, number
of correct responses, etc. Each task model would continue to be
built on a lesson-by-lesson basis, depending on the structure and
needs of each lesson.
[0034] The `Assessment Results` module contains the results of any
assessment instruments applied outside the context of a lesson (any
kind of pre-tests or post-tests, for example).
[0035] Finally, the `Student Concept Map` module contains a
partially ordered set of "concept nodes", where each concept node
represents one of the abstract concepts that the lessons teach to
the student. Each concept node contains a numeric rating of the
student's mastery of that concept, and references to lesson entries
in the Lesson Task Models/Results module which provide evidence for
those ratings. One goal of the AI Engine is to be able to predict
values for some of the information in the Student Cognitive Model.
All information that can be predicted is tagged with a certainty
rating. A certainty rating of -1 means no information is known
about the data, while a certainty rating between 0 and 1 means some
information is known (with 0 meaning the value is based purely on a
predictive model and 1 meaning that the value is known as fact).
For example, the AI Engine might predict that a student is a visual
learner. At this point, the student's learning style would
initially have a certainty value of 0. As the student works within
the learning system, the system gathers additional evidence (the
student's performance on lessons designed for a visual learning
style) to confirm or modify that predictive classification. The
more consistent evidence the system gathers, the higher the
certainty rating becomes.
[0036] FIG. 3 illustrates how the Student Cognitive Model resides
at the core of the present invention learning system. By being a
repository of information about the student, what the already know,
and his or her progress within the learning environment, the
Student Cognitive Model facilitates asynchronous communication
between the predictive AI Engine, the Agents, and the student's
parents and/or teachers. The Student Cognitive Model incorporates
the currently known cognitive state of the student. For example, if
the student's prior knowledge information indicates that the
student is interested in sports, then activities that relate to
sports and games are preferred. Table 1 illustrates the manner in
which each of these entities makes use of and/or contribute to the
Student Cognitive Model.
[0037] FIGS. 4 and 5 show the Agent architecture and how it
interfaces to the learning environment through an Application
Programming Interface (API). In this architecture, the agent is
very modular and is not tightly integrated with the learning
environment at all--in fact, it only communicates with the learning
environment through a well-defined interface. It is built around
the Agent Control Module, which is responsible for processing
events received from the learning environment, updating its student
and session models based on those events, initiating interactions
with the student, and potentially sending commands back to the
learning environment. The appearance and user interface (UI)
aspects of the agent have been separated out into a distinct
module. This allows for maximum flexibility in the appearance of
the agent. Either the agent can appear in its own window, or the
learning environment itself can implement this module, allowing the
agent to appear within the main window of the environment, as is
the case with the present invention agents. This architecture is
very flexible as shown in being able to relax the architectural
separation between the agent and the learning environment. In this
way the Content Agents are not required to be tied to those lessons
and tightly integrated with them. However, even if the Agents are
implemented in the same code as the lessons themselves, it is still
useful to separate them on a conceptual level. Relaxing this
architectural separation simplifies some of the Agent's work, as
the Agent then has complete access to the state of the learning
environment, rather than getting only the information disclosed
through the Agent/environment API.
[0038] Domain knowledge is also tied directly to each Content
Agent, and there is duplication of domain knowledge across the
various Content Agents. For this reason there is also a knowledge
base of domain knowledge separate from any one particular
Agent.
[0039] Inside the Agent Behavioral Control Description, there are
four categories of information--pedagogical goals, potential
student actions, agent interventions, and the student concept
database. `Pedagogical goals` represent the general educational
objectives of the Agent--for example, what concepts are most
important for a student to learn, what learning style does the
student prefers. `Potential Student Actions` are exactly
that--actions or (more often) classes of actions that the student
may take within the learning environment. `Agent Interventions`
comprise all actions that the agent might take, such as giving the
student a clue, explaining a concept, or adjusting the lesson
content or goals.
[0040] The `Student Concept Database` is the generic version of the
student concept map we have already been discussing in the context
of the Student Cognitive Model. It contains the concepts and their
relationships, with none of the data regarding where a particular
student is at in understanding the concepts.
[0041] The relationships between these four groups of entities in
the Agent Behavioral Control Description are expressed in terms of
declarative rules, such as, if the student performs potential
student action X it is evidence that they have a misconception
regarding concept Y and the Agent should take intervention Z.
[0042] The Agents base their actions on a Student Cognitive Model
and task models that they construct via their interactions with the
student and by observing the student's performance on multiple
tasks in the learning environment. A Student Cognitive Model is a
model of the student's cognitive state, relative to the educational
subject area. It may also contain other information about the
student, such as preferences, learning style, etc. Task models are
generally somewhat simpler--they model the particular educational
task the student is performing and how the student is progressing
through that task.
[0043] The present invention learning system contains software
Agents. The "Content" characters in each lesson can be seen as
tutors or instructional Agents, while the Helper character
personifies an Agent designed to shape and guide the student's
overall learning experience, and to provide positive feedback
response to the student. From an affective perspective, the Helper
Agent also acts as the student's companion and "buddy" throughout
the learning experience. The Helper Agent is in charge of lesson
selection and providing assistance and encouragement to the
student, while the Content Agents in each lesson are in charge of
managing the presentation of the subject matter.
[0044] The drawing shown in FIG. 6 clearly identifies the two
distinct modalities for the employment of AI techniques within the
present invention. The first, identified in the figure as the "AI
Engine" component, employs offline machine learning designed to use
the Student Cognitive Model of existing students to automatically
learn classifiers which can then be used to classify new students
according to their initial profiles along such axes as preferred
stories, music, learning styles, attention spans, etc. This initial
classification can then be used to provide the first level of
lesson selection and customization for the new student. The second
component, identified in the figure as "Student Cognitive Model",
shows the use of an interactive pedagogical Agent to monitor the
student's progress and work on the student's behalf to provide
customized guidance and lesson planning, based on the ongoing
cognitive model and actual student performance within the learning
system.
AI Engine
[0045] The AI Engine component of the present invention takes as
input the ongoing cognitive models of existing students working
through the system patterned against the cognitive model of new
students. Its output is a classification system that is used to
predict the best lessons to assign to the new students, given only
their entrance cognitive state. Let's take a look at one example of
how this works:
[0046] Suppose that a students profile shows that they have prior
knowledge from watching movies, in particularly the Harry Potter
movies. The system can now predict that questions about the movie
can be answered with high success, and that presented material in
the context of the movie will fit the cognitive model of the child.
Using this information the student can be classified as one that
benefits from storylines. By pairing the student's entrance
cognitive model with this outcome a training example is made. This
training example, along with the training example from other
students, would be used by the AI Engine to product a general
classifier.
[0047] As a second example suppose that a student did very well in
lessons that were designed specifically for a visual learning
style, but was less successful in similar lessons designed for
other learning styles. We might classify this student as a
primarily visual learner. This student's entrance cognitive state
would then be paired with that classification to make a "training
example". That training example, along with the training examples
from all other existing students in the system would then be
provided to the AI Engine to produce an automated "classifier". The
classifier in this example would be a mapping from entrance
cognitive models to learning styles. Once it has been created by
the AI Engine, we give it the entrance cognitive model of a new
student and it will predict what their learning style is.
Generally, the more training examples we have when building the
classifier, the more accurate the resulting classifier will be. To
understand how the classifier is built we can see that it depends
upon what machine learning algorithm is used and the specific
relationships revealed from the data.
[0048] Machine learning techniques used in the present invention
are known as "supervised" machine learning. The learning is said to
be supervised because we, the "teachers", are providing the system
with a set of pre-classified training examples. This type of
learning is also known as "inductive reasoning", because the system
knows no logical rules about how to classify students until it is
given the training examples that it analyzes to "induce" a
classification system. The data that we send to the classifier (in
our case, a Student Cognitive Model) is known as an "input
vector".
Machine Learning AIgorithms
[0049] While "Decision tree algorithms" that analyze the training
data they are given to produce a tree-structured classifier, and
"Case-Based Reasoning algorithms," that look for the "closest"
previously seen example in a database and output the classification
from that example, were considered for the present invention, an
Artificial Neural Network (ANN) was chosen as the most appropriate
machine-learning algorithm. It incorporates decision tree
algorithms and case-based reasoning algorithms along with other
logic-based systems in a complex network similar to human neural
functions.
Knowledge And Skill Domains For Instruction
[0050] The framework of the present invention learning system
curriculum consists of specific instructional knowledge and skill
areas derived from the research literature (see FIG. 9). These
areas describe, at a high level, student proficiency targets,
taking into account age/grade appropriate content. Examples of
these skill areas include the five areas of proficient reading
identified by the National Reading Panel including phonemic
awareness, phonics, vocabulary, fluency, and comprehension. The
six-traits of effective writing provide a strong foundation for
organizing writing instruction and collectively define writing
proficiency. The writing traits include ideas, organization, voice,
word choice, sentence fluency, and conventions. In mathematics, the
National Teachers of Mathematics have defined ten school content
areas including computation, algebra, geometry, measurement, data
analysis and probability, problem solving, reasoning and proof,
communication, connectionns, and representation. These areas are
reflected at all levels of the proposed mathematics curriculum
applied to grade and age-level appropriate content.
[0051] These knowledge and skill areas in the domains of reading,
writing, and mathematics provide a framework for organizing and
sequencing the instructional curriculum. Artificial intelligence
sub-systems and Agents flexibly move students through the
curriculum, considering the skill gaps, mastery retention,
preferred learning styles, and other performance and demographic
predictors of domain and content area development needs.
Critical Scope And Sequence
[0052] The scope and sequence of the present invention curriculum
follows the instructional scaffolding of content area knowledge and
skills, building increasingly complex ability. Throughout the
curriculum, basic skills are introduced, taught explicitly,
practiced and reinforced. Skill level performance is distributed
throughout the curriculum in order to monitor student confidence
and skill retention. The curriculum was developed with the
following points of guidance:
[0053] Appropriate grade or age-level content at each level of the
curriculum
[0054] The knowledge and skill order within each level of the
curriculum
[0055] Curriculum scaffolding for introducing, explicitly teaching,
practicing and retaining knowledge and skill mastery
[0056] The AI Engine and the actions of the Agents drive the
present invention curriculum. These systems use an array of student
demographic, learning preference, and performance data, shown in
FIG. 7, to make strategic decisions about the order and nature of
instructional lessons to provide to each student. Personalization
of the curriculum is achieved, in part, through the actions of
these systems.
[0057] Students' demographic data include, for example, the
student's age, enrolled grade, prior reading level, diagnosed
special needs, family structure, income level, and type of
community and school. Data about students' instructional
preferences include their preferred learning style(s) and
instructional structure. Further data includes information about
the existing knowledge base of the child to support a cognitive
model. These data are obtained from initial parent and teacher
feedback, questionnaires, tests, preferences, hobbies, and
experiences. As students interact with lessons, preference data are
collected from the type of lessons the student easily masters or
struggles to master. The teaching styles and structures, inherent
in these lessons, are used to update the preference data for the
students.
[0058] Student performance data are obtained from the assessment
system that includes three levels of tests: The Initial Placement
Test, Instructionally Imbedded Tests and Benchmark Tests, as shown
in FIG. 8.
[0059] The Initial Placement Test provides placement of students
into the curriculum and establishes a baseline for judging one-year
growth in knowledge and skill performance. The Instructionally
Imbedded tests provide information for guiding students through the
curriculum at an appropriate rate, assuring mastery of foundation
skills, and providing information for reporting student's progress.
The Benchmark Tests provide post-test information for comparison
with the initial placement information and the retention of mastery
of knowledge and skills from the skill lessons.
[0060] The curriculum is taught in a flexible, sequential teaching
strategy that allows students to move at their own developmental
pace through the lessons. They advance ahead or repeat lessons
based on their performance within the lessons. The assessment
system is designed to support the flexible, sequential curriculum.
The AI Engine and the Agents drive the optimal selection of lessons
to obtain mastery of the desired knowledge and skills. Skill gaps
and instructional sequences are determined for each child based on
the instructional needs of the child and their learning preferences
as deduced from the Student Cognitive Model.
Initial placement test
[0061] The initial placement test provides data to help place
students in the appropriate levels of the curriculum and to
identify the skill gaps that need to be addressed. To accomplish
this and to keep the length of the assessment as short as possible,
the assessment are constructed in two parts, a general placement
and a specific knowledge skills assessment associated with the
general placement. The general placement of each student is used to
focus the second skills part of the test.
[0062] The AI Engine and Agent systems monitor the child's
performance on the initial lessons to insure the appropriate
placement of students in the curriculum. Students may need to be
placed earlier in the instructional sequence if they are struggling
with the first lesson. If they are doing well, they may need to
advance to the next lesson. This monitoring helps further refine
the initial placement of students in the curriculum.
Embedded tests
[0063] The skill lessons at each grade level include multiple
learning tasks (game-like responses from students). These responses
are monitored to determine the students' mastery of the lesson
content. The performance of the students is used to determine when
students move to the next lessons or when they receive additional
instruction and skill practice. Where possible, the skill errors
students make are tracked and specific practice and instruction is
offered focusing on these errors.
Benchmark tests
[0064] Among the curriculum lessons, periodic benchmark tests are
provided to assess the retention of critical scaffolding skills.
These tests are used to identify persistent skill gaps and to
assure that prerequisite skills are mastered prior to the
instruction of higher-order skills. The benchmark tests include
knowledge and skills that have been previously taught. Performance
on the benchmark tests are compared and reported to document the
learning gains attributable to instruction.
Critical Teaching Strategies
[0065] The instructional methods included in the curriculum
generally follow the behavioral strategies suggested by Hunter
(1991). The curriculum is built from a set of reading standards
that reflects state and national standards, stated as behavioral
knowledge and skills suggested by the research literature. The
explicit instruction of skills includes modeling, guided practice,
assessment, and extended practice if required for student mastery
of the skills.
[0066] Three general teaching models are implemented in the
Curriculum:
[0067] Skill-based instruction (modeling, practicing,
assessing)
[0068] Task complexity instruction (sequence of increasingly
complex tasks)
[0069] Project instruction (planning, execution, review,
revision)
[0070] Skill acquisition is distributed throughout the curriculum.
Pre-skills are introduced prior to explicit skill instruction and
are reinforced through applications in new skill development
activities. The specific instructional sequences for each level of
the curriculum are outlined in the attachments, along with the
specific curriculum standards.
[0071] The teaching strategies are guided by the AI Engine and the
Agents to assure that the teaching methods contribute to the
success of the students. Where possible, the student's preferences
are employed to reinforce mastery of the specific lessons content.
Recursive instruction, to address skill gaps, is used and
modifications of instructional strategies are included in the
decision rules used to guide such repetition. The AI Engine and the
Agent systems are critical elements of the success of the
instructional system.
Specific Embodiments
[0072] One embodiment of this invention is a system that delivers
instruction to computers over a digital network such as the
Internet. This includes instruction delivery to computers and
handheld devices via wire and wireless means such as Ethernet and
WiFi. This embodiment does not preclude the use of this invention
in other networked and non-networked digital environments, nor
delivery to alternative digital devices.
[0073] In this embodiment, functionality is implemented on both the
student's computer or `Client` and on the server computer or
`Server`. The server contains all program code and data stored in
permanent disk memory and program memory. The Server transfers
program code and data to the Client as needed.
[0074] In the present embodiment the major components of the Server
consist of a AI Engine, Learning Management System (LMS),
Activities and Tests, Agents, web server, database storage, and the
state of cognitive models.
[0075] The Client receives appropriate instructional material over
the network for delivery to the student and sends results of the
student's interaction with this material back to the Server.
[0076] The Learning Management System (LMS) authenticates students
onto the system, assigns instructional lessons to them (based on
information from the AI Engine), measures and saves responses,
ensures the sequence of instruction is presented and completed
correctly, and provides administrative reports of student progress.
The LMS is also a Content Management System in that administrators
may add or remove lesson content to and from the system in an easy
way while the system provides revision and access control. The LMS
also provides the User Interface or UI of lessons to the
Student.
[0077] The Instructional Content is contained in Macromedia
Flash-based `Activities` which consist of 1 to several `Lessons`.
Each Activity teaches a specific area of study while each Lesson
focuses on different areas within this study area. The Flash
Activities are uploaded to the Client computer where they are
executed using a local Flash `Player` This reduces the load on the
server since the resource heavy Activities, featuring animations
and complex interactions, are run entirely on the Client, the
Server simply has to upload the Flash Device as a binary file to
the Client. While the Client executes the Lessons in an Activity
however, each Lesson communicates back to the Server about progress
and other result events.
[0078] The state of the Agents is always maintained on the Server
even though oftentimes they are executing on the Client. Each
instructional Activity has its own Content Agent to interface
between the student and the functionality of the Activity, and to
also interface with the Learning Management System. The Content
Agents are in fact experts about the instructional material of
their particular Activity and contribute to the corresponding
Student Cognitive Model components. In addition, the Helper Agent
manages all of the Content Agents. It is the Helper Agent who
immediately controls the system in not only managing Agent-input to
the Student Cognitive Model but also ongoing input to the Student
Cognitive Model from the AI Engine.
[0079] The AI Engine is a large component of the web server system
and is a dedicated server in its own right. The AI requirements of
the system are computational heavy and the dataset is large. For
this reason the classifier output of AI Engine is only generated
infrequently. Existing classifications are immediately available to
the system to apply to new students. The information from the AI
Engine is used by the LMS to initially assign lessons to
students.
Operation
[0080] As shown in FIG. 10, a single Server can service multiple
"Clients". The Clients are the end-user student computers located
at home or at an institution such as a school. The Server(s) is
located as the company's facilities or an Independent Service
Provider (ISP). The Server works in the background and is normally
not physically seen by the student.
[0081] The parent or guardian of the student first browses to the
online purchase area of the Server and during this process enters
background information about the student. This information is used
to populate the prior knowledge Student Cognitive Model of the
student. Teachers may also enter information in this location. The
information gathered includes the interests of the child,
experiences they have had--games they play, movies they've watched,
stories they are familiar with, people they know, television
programs that they watch, their likes and dislike. This information
also includes information about the family such as native language,
learning issues with other family members, and educational history.
Direct information about the child such as age, grade, sex, spoken
language, currently estimated education level, and
dyslexia-screening question results, are also entered at this
point. All of this information is used to populate the cognitive
state of the Student Cognitive Model.
[0082] The first experience the student has with the system is to
identify further interests that they personally relate to. A
dynamically adaptable test is also given to the student that
directly assesses their current competence in the targeted area of
study such as reading, mathematics, and science. This information
is also used to populate the cognitive components of the Student
Cognitive Model. The parent/teacher also receives an online
assessment of the results of the student test, and further
recommendations, should issues such as the potential for dyslexia
be identified.
[0083] The Server side AI Engine now patterns the new student
cognitive model against the cognitive models of all existing
students to predict the best course of study for that particular
student. Using an AI Engine means that perfect matches do not need
to be made to the extent that the AI Engine may generate a new
course combination that has never been recommended before. This
customized program of study is then sent to the parent/teacher or
guardian and the student may begin the program.
[0084] An important component of the customized program of study
are the Agents. These software Agents provide ongoing help and
guidance to the student and have a visual component that appears as
animated characters. The Agents have access to the Student
Cognitive Model as initially communicated from the Server-side AI
Engine when the customized program for the student is first
acquired by the parent/teacher or guardian. There are two main
classes of Agents: the "Helper" Agent, and the "Content" Agents.
The Helper Agent acts as a single guide and accompanies the student
throughout the entire program by recommending paths, making
introductions to other Agents, taking a break from the program,
answering questions, providing encouraging feedback, and acting as
a companion to the student. The Helper Agent is not a content
domain expert in themselves but they know where the content
information is located. The Helper Agent may contain a voice
recognition system where the student may speak questions and
instructions to the Agent. The Agent mostly communicates to the
student by visual and auditory means.
[0085] The "Content" Agents are located at each different area of
content, or skill area, know as an Activity. It is the Helper Agent
that navigates the student to the various content activities where
the Content Agents reside. Unlike the Helper Agent, the Content
Agents are experts in their subject domain. These Agents are aware
of the predicted course of study for the student and are constantly
evaluating this predication based on student responses to the
lesson material. Should student responses align with the prediction
then the predication value is strengthened, should the student
responses not align with the prediction then it is modified based
on the Student Cognitive Model. The Content Agent introduces the
subject matter, instructs the student on required background
knowledge, demonstrates how the student should operate the
game-like Activity, and acts as an expert to answer questions or
offer help when the student requires it. The Content Agents
communicate with one another and the Helper Agent through
information in the Student Cognitive Model.
[0086] The content Activities are fun, highly animated game-like
exercises that focus on a particular skill, or area of study. These
Activities are also adaptive in that they monitor how the student
progresses and make adjustments accordingly. This information is
communicated to the Helper Agent who then updates the Student
Cognitive Model with new information about the Student. The Helper
Agent also takes action based on the newly updated cognitive model.
For example, if the student struggles in a particular Activity then
the Activity itself will first implement internal changes to adapt
to the student. If the problem goes beyond the scope of the
Activity then the Helper Agent, in conjunction with the Student
Cognitive Model, will adapt at the Activity level to redirect the
student to less difficult material, revisit an Activity with new
material, or direct the student to a similar Activity using an
alternative learning style for example. Likewise if the student
does particularly well in an Activity outside of its scope then the
Helper Agent may direct the Student to more challenging activities.
It is important to note that the Helper Agent can draw on
information in the Study Cognitive Model, which in turn has been
initially predicted and created by the AI Engine. For example,
student progress might indicate that the student is mastering the
material and that the child should be moved to more challenging
material. However, upon consulting the Student Cognitive Model the
Helper Agent may discover that information patterned from the AI
Engine indicates that past experience has shown that this
particular student is probably too young to warrant this change, so
an alternative activity, with perhaps more focus on a different
area of study is given.
[0087] At this point it can be seen how the Server side AI Engine
initially predicts a best course of study for the student based on
an AI pattern of their cognitive model with cognitive models of a
previous population of students. The AI engine then recommends an
initial program that it has found to be historically the most
successful for this Student Cognitive Model. Once the student
begins the program the Agents then take over the role of
dynamically refining the cognitive model and adapting the program
to the needs of the student. In addition to information about the
student's progress and responses received from the Activities to
the Helper Agent, the Activities themselves also contain specific
embedded tests to assess how well the child is doing. This further
helps to refine the Student Cognitive Model on an ongoing
basis.
[0088] As the student progresses though the program the Agents
further refine the Student Cognitive Model that in turn continually
adapts the content material to the student. By using the program
this way the system "learns" more and more about the best
techniques and strategies that work with each individual student.
This information is also communicated back to the AI Engine where
it is added to the pool of past student population histories thus
further refining the model for future students. As more and more
students pass though the system it can more accurately predict what
the best initial program of study is, and what the best action to
take on a going basis should be. The results are that the student
learns in an environment that builds on the knowledge they already
have by integrating new information with this knowledge. These
connections are then strengthened in specifically designed
Activities to ensure that real learning has occurred. Similarly,
other Activities measure the ability of the student to recall this
information in a meaningful and useful way so that the new
information becomes integrated into what the student already knows.
It is at this stage the student has "learned" the new
information.
* * * * *